Scholarlink

<p>Dermoscopy improves early detection of pigmented and non-pigmented skin cancers, yet interpretation remains operator dependent and variable. Artificial intelligence (AI) systems trained on dermoscopic images offer scalable decision support, but evidence is dispersed across datasets, model types, and reference standards, limiting clinical translation. This systematic review aimed to evaluate the diagnostic accuracy of artificial intelligence (AI)-based algorithms for dermoscopic image analysis across various skin conditions and to compare different models based on dataset characteristics, training strategies, and diagnostic reference standards. We systematically searched PubMed, Scopus, Cochrane Library, and Web of Science up to May 2025 using comprehensive search strategies. Studies were included if they applied AI techniques to dermoscopic images for diagnostic purposes and reported performance metrics. Risk of bias was assessed using the QUADAS-2 tool. Data were synthesized qualitatively due to methodological heterogeneity. Ninety-six studies met the inclusion criteria, covering a wide range of AI approaches. Convolutional Neural Networks (CNNs) were the most common model type and generally showed variable diagnostic performance, with sensitivity ranging from 0.69 to 1.00 and specificity from 0.36 to 0.98. Hybrid and ensemble models, especially those incorporating EfficientNet, DenseNet, and ResNet architectures, often outperformed classical machine learning approaches. Key limitations included variability in datasets, inconsistent ground truth references, and lack of standardized reporting. Few studies involved prospective real-world validation. AI-based algorithms for dermoscopic image analysis particularly CNNs and ensemble frameworks, have diagnostic potential, however, variability in study design and evaluation limits their current clinical applicability. Future efforts must focus on standardization, external validation, and transparent algorithm development to support clinical integration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative evaluation of AI algorithms for dermoscopic image analysis in skin disease diagnosis: a systematic review of model accuracy and methodological heterogeneity

  • 期刊论文

  • 发表日期：

    2026年1月30日

查看全文
  • Areeba Ahmed,
  • Meredith Hengy,
  • Steven Daveluy

摘要

Dermoscopy improves early detection of pigmented and non-pigmented skin cancers, yet interpretation remains operator dependent and variable. Artificial intelligence (AI) systems trained on dermoscopic images offer scalable decision support, but evidence is dispersed across datasets, model types, and reference standards, limiting clinical translation. This systematic review aimed to evaluate the diagnostic accuracy of artificial intelligence (AI)-based algorithms for dermoscopic image analysis across various skin conditions and to compare different models based on dataset characteristics, training strategies, and diagnostic reference standards. We systematically searched PubMed, Scopus, Cochrane Library, and Web of Science up to May 2025 using comprehensive search strategies. Studies were included if they applied AI techniques to dermoscopic images for diagnostic purposes and reported performance metrics. Risk of bias was assessed using the QUADAS-2 tool. Data were synthesized qualitatively due to methodological heterogeneity. Ninety-six studies met the inclusion criteria, covering a wide range of AI approaches. Convolutional Neural Networks (CNNs) were the most common model type and generally showed variable diagnostic performance, with sensitivity ranging from 0.69 to 1.00 and specificity from 0.36 to 0.98. Hybrid and ensemble models, especially those incorporating EfficientNet, DenseNet, and ResNet architectures, often outperformed classical machine learning approaches. Key limitations included variability in datasets, inconsistent ground truth references, and lack of standardized reporting. Few studies involved prospective real-world validation. AI-based algorithms for dermoscopic image analysis particularly CNNs and ensemble frameworks, have diagnostic potential, however, variability in study design and evaluation limits their current clinical applicability. Future efforts must focus on standardization, external validation, and transparent algorithm development to support clinical integration.